OpenMind

# OpenMind — Distributed AI Compute Mesh <p align="center"> <img src="OpenMind.png" alt="OpenMind Logo" width="200" /> </p> **Pool your laptops, desktops, and external GPUs into a private AI cluster.** Chat with large language models sharded across every machine. Generate images and videos using every GPU in the network. All self-hosted, encrypted, peer-to-peer. Built by integrating three open-source projects into one unified system, branded as **OpenMind**: | Project | Role | |---------|------| | **[Odysseus](https://github.com/pewdiepie-archdaemon/odysseus)** (66.8k ⭐) | Per-person AI workspace — chat, agents, docs, memory, email, calendar | | **[Hyperspace AGI](https://github.com/hyperspaceai/agi)** (1.9k ⭐) | P2P distributed LLM inference with model sharding | | **[ComfyUI-Distributed](https://github.com/robertvoy/ComfyUI-Distributed)** (564 ⭐) | Multi-GPU image & video generation across machines | --- ## Architecture ``` OpenMind (your workspace — Odysseus under the hood) │ ├─→ pod_manager MCP ──→ Hyperspace Pod (:8080) ──→ LLM inference (sharded across GPUs) │ ├─→ video_gen MCP ──→ Dispatcher (:9000) ──→ ComfyUI Master ──→ Workers (parallel video gen) │ └─→ Tailscale mesh (encrypted, 100.x.x.x) ``` - **Chat**: Hyperspace shards models across pod GPUs for instant responses - **Video**: ComfyUI-Distributed splits frames/batches across workers - **Image**: Same pipeline, same workers, different models - **Workspace**: Full Odysseus feature set — agents, deep research, email, calendar, memory, document editor, cookbook --- ## Quick Start ### Prerequisites - 2+ machines with GPUs (at least one with 8GB+ VRAM for video) - Tailscale account (free for up to 3 users) - Docker (optional, for containerized deployment) ### 1. Network Layer (all machines) ```bash # Install Tailscale curl -fsSL https://tailscale.com/install.sh | sh sudo tailscale up # Note each machine's 100.x.x.x IP tailscale ip -4 ``` ### 2. Hyperspace Pod (all machines) ```bash # Install Hyperspace CLI curl -fsSL https://agents.hyper.space/api/install | bash # On coordinator: hyperspace pod create "openmind-cluster" hyperspace pod invite --role member --ttl 168h # → Share the invite code with pod members # On workers: hyperspace pod join <invite-code> # Shard a model across all GPUs: hyperspace pod shard qwen3.5:32b # Test inference: hyperspace pod infer -p "Hello from OpenMind!" ``` ### 3. ComfyUI Cluster ```bash # Clone ComfyUI on each GPU machine git clone https://github.com/comfyanonymous/ComfyUI.git ~/comfyui # Install ComfyUI-Distributed extension git clone https://github.com/robertvoy/ComfyUI-Distributed.git \ ~/comfyui/custom_nodes/ComfyUI-Distributed # Start master (coordinator): cd ~/comfyui && python main.py --enable-cors-header --listen # Start worker (other machines): cd ~/comfyui && python main.py --enable-cors-header --listen # Configure worker in ComfyUI UI → Distributed panel → enter master's Tailscale IP ``` ### 4. Dispatcher (coordinator only) ```bash pip install fastapi uvicorn httpx pyyaml pydantic # Edit dispatcher/config.yaml with your IPs, then: python dispatcher/main.py ``` ### 5. OpenMind MCP Integration Copy the new MCP servers into Odysseus: ```bash cp mcp_servers/video_gen_server.py ~/odysseus/mcp_servers/ cp mcp_servers/pod_manager_server.py ~/odysseus/mcp_servers/ ``` Add to Odysseus config: ```yaml mcp_servers: video_gen: command: python args: [mcp_servers/video_gen_server.py] pod_manager: command: python args: [mcp_servers/pod_manager_server.py] ``` --- ## One-Command Full Setup ```bash chmod +x setup.sh ./setup.sh # auto-detect role ./setup.sh --role coordinator # force coordinator ./setup.sh --role worker # force worker ``` --- ## What You Can Do in OpenMind ### Chat ``` "Summarize this 50-page PDF" → Routed to Qwen 3.5 32B sharded across 3 GPUs ``` ### Video Generation ``` "Create a 10s cinematic of a dragon flying over mountains at sunset" → Routed to ComfyUI worker with Wan 2.2 (24GB GPU) → Parallel variations generated across multiple workers ``` ### Image Generation ``` "Generate a realistic photo of a cyberpunk city street at night" → Routed to ComfyUI worker with SDXL/Flux → 4 variations in the time of 1 ``` ### Pod Management ``` "Show me who's online and what GPUs are free" → pod_manager MCP returns live pod status ``` ### All Standard Odysseus Features Chat · Agents (bash/files/web) · Deep Research · Document Editor · Memory (ChromaDB) · Email (IMAP/SMTP) · Calendar (CalDAV) · Model Comparison · Cookbook (hardware scan + model recommendations) · 10,000+ MCP tools · PWA mobile access --- ## Model Reference ### LLM (shardable across pod) | Pod VRAM | Model | |----------|-------| | 16 GB (2×8) | Gemma 3 12B, Qwen 2.5 14B | | 32 GB (2×16) | Qwen 3.5 32B | | 64 GB (4×16) | Qwen 2.5 72B (Q4), Llama 3.1 70B (Q4) | | 96 GB+ | DeepSeek V3 (Q4) | ### Video (per single GPU — not shardable) | GPU VRAM | Model | |----------|-------| | 8-12 GB | LTX-Video, AnimateDiff | | 16-24 GB | Wan 2.2, CogVideoX-5B | | 24-48 GB | Wan 2.2 (1080p), Hunyuan Video (Q8) | | 48-80 GB | Hunyuan Video (full), CogVideoX-5B-I2V | --- ## File Layout ``` openmind/ ├── OpenMind.png ← Logo ├── README.md ← This file ├── ARCHITECTURE.md ← Full architecture documentation ├── docker-compose.yml ← Full stack Docker deployment ├── setup.sh ← One-command automated setup ├── dispatcher/ ← Glue service (FastAPI) │ ├── main.py ← Central job router │ ├── config.yaml ← Configuration │ ├── requirements.txt │ └── Dockerfile ├── mcp_servers/ ← MCP tools (drop into Odysseus) │ ├── video_gen_server.py ← Video generation │ └── pod_manager_server.py ← Pod management + chat └── comfyui_workflows/ ← Pre-built ComfyUI workflows ├── wan22_t2v.json ← Wan 2.2 text-to-video ├── ltx_video.json ← LTX-Video ├── hunyuan_video.json ← Hunyuan Video └── sdxl_image.json ← SDXL image generation ``` --- ## FAQ **Q: Does OpenMind replace Odysseus?** No — OpenMind augments it. You still run Odysseus as your workspace. OpenMind adds two new MCP servers (video generation + pod management) and the Dispatcher/ComfyUI infrastructure to pool pod GPUs. All existing Odysseus features (chat, agents, memory, email, calendar, cookbook, deep research) work as before. **Q: Can I use this with just CPU machines?** LLM inference via Hyperspace will work but slowly. Video generation requires at least one GPU. For CPU-only, configure a cloud provider as fallback (`hyperspace pod providers add`). **Q: What if the coordinator goes offline?** Hyperspace Pod (LLM) is fully P2P — no single point of failure. ComfyUI-Distributed has a master-worker model — if the master dies, video gen stops. Run the master on the most reliable machine. **Q: Is this actually private?** Yes. Tailscale encrypts all traffic with WireGuard. Hyperspace uses Noise-encrypted libp2p. ComfyUI workers bind to Tailscale IPs only. Nothing is exposed to the public internet. **Q: What's the minimum VRAM for video?** LTX-Video runs on 12GB. For a group with mixed hardware, start with LTX-Video as the baseline. --- ## 🏗️ Built On OpenMind would not exist without these three incredible open-source projects. All credit for the core engines goes to their respective teams: <table> <tr> <td align="center" width="33%"> <a href="https://github.com/pewdiepie-archdaemon/odysseus"> <img src="https://img.shields.io/github/stars/pewdiepie-archdaemon/odysseus?style=for-the-badge&logo=github&label=Odysseus" alt="Odysseus Stars" /> </a> <br /> <b>Odysseus</b> <br /> Per-person AI workspace — chat, agents, deep research, docs, memory, email, calendar <br /> _{License: Apache 2.0} </td> <td align="center" width="33%"> <a href="https://github.com/hyperspaceai/agi"> <img src="https://img.shields.io/github/stars/hyperspaceai/agi?style=for-the-badge&logo=github&label=Hyperspace%20AGI" alt="Hyperspace Stars" /> </a> <br /> <b>Hyperspace AGI</b> <br /> P2P distributed LLM inference with model sharding via libp2p <br /> _{License: MIT} </td> <td align="center" width="33%"> <a href="https://github.com/robertvoy/ComfyUI-Distributed"> <img src="https://img.shields.io/github/stars/robertvoy/ComfyUI-Distributed?style=for-the-badge&logo=github&label=ComfyUI-Distributed" alt="ComfyUI-Distributed Stars" /> </a> <br /> <b>ComfyUI-Distributed</b> <br /> Multi-GPU image & video generation across machines <br /> _{License: MIT} </td> </tr> </table> > **OpenMind is glue code.** The real magic comes from these three projects. Please ⭐ star them! --- ## License MIT — all glue code in this repo. Individual components retain their original licenses (Odysseus: Apache 2.0, Hyperspace: MIT, ComfyUI-Distributed: MIT).